Enhancing Autonomous Decision-Making Through Ontology-Driven Natural Language Understanding

Autonomous systems highly rely on natural language understanding (NLU) to process human instructions to make intelligent decisions in dynamical environments. However, classical models of natural language processing subjected to weaknesses of contextual semantics and domain-specific knowledge, result in ambiguous interpretations and lower accuracy of decision making. To overcome this issue, in this paper, an ontology-based system is derived to improve autonomous decision-making using novel Bigram Knowledge Lightweight Deep Neural Network (Bigram-KL-DNN) model. The proposed Bigram-KL-DNN method integrates the bigram based contextual feature extraction and ontology based guided semantic representation to enhance the meaning of the natural language instructions and queries. Bigram features within the framework capture contextual relationships between successive words, whilst ontology knowledge gives the structured semantic relationships between the concepts within the domain of the user intent. Those extracted features are fed into a lightweight deep neural network classifier are trained to produce high accuracy of classification with less computing complexity. The proposed model is assessed based on a dataset of 50,000 natural language samples that are divided into five independent decision classes, such as device control commands, environmental monitoring, security verification, healthcare assistance, and general assistance queries. Experiment shows that a proposed Bigram-KL-DNN model attains an average of 97.39, precision of 96.98, recall of 96.90 and F1-score of 96.93 as compared to several available models like traditional N-gram model, CNN-based models, LSTM semantic model and transformer-based architecture. Additionally, the proposed model is more computationally efficient with a median processing time of 28 ms and a reduced memory utilization of 410 MB and is suitable in autonomous systems that work in real-time. The scalability and reliability of the proposed framework is also confirmed by robustness analysis that was conducted using large datasets.